MOSS-Image Four-Step Distillation LoRA

Community author lvladikov released a 4-step distillation LoRA (rank 64) for OpenMOSS's MOSS-Image SFT training, uploaded on 11 19, with the current checkpoint produced on 11 20. The method uses eDMD distribution matching, aligning with a teacher model of 50 steps and CFG 3, plus a small LADD-style adversarial discriminator that compares student images against teacher images and approximately 8,000 cropped real photos. Image generation requires only 4 steps and CFG 1, with one model call per step (the teacher uses 100 calls in total); training took approximately 57 hours on a single NVIDIA B300, covering 12 resolutions ranging from 768×512 to 1824×1024.

Original title: lvladikov/MOSS-SFT-Distill-4step-LoRA

Community author lvladikov released a 4-step distillation LoRA (rank 64) for OpenMOSS's MOSS-Image SFT training, uploaded on 11 19, with the current checkpoint produced on 11 20. The method uses eDMD distribution matching, aligning with a teacher model of 50 steps and CFG 3, plus a small LADD-style adversarial discriminator that compares student images against teacher images and approximately 8,000 cropped real photos. Image generation requires only 4 steps and CFG 1, with one model call per step (the teacher uses 100 calls in total); training took approximately 57 hours on a single NVIDIA B300, covering 12 resolutions ranging from 768×512 to 1824×1024.

Why it matters The author claims image generation GPU time is reduced by approximately fourteen times, with color tones close to the original model and text largely rendered correctly; it can run using only ComfyUI core nodes

Upvotes 30 · Likes 33 · Downloads 2685

Reasons

  • The author claims that on Comfy Cloud, for the same 22 images of size 1024×1024, billed GPU time is 26.6 seconds versus 387.3 seconds, saving approximately 14 times; across 21 test prompts, average brightness is 1.0 L* lower than the teacher, and fine texture detail is 1.16 times that of the teacher
  • Of 15 text prompts, 14 rendered all required characters correctly (the teacher rendered 15, and akhaliq's 4-step LoRA rendered 12)
  • Includes ten ComfyUI workflows (text-to-image, layout, image editing, PiD 4x upscaling, etc., with one local and one Comfy Cloud version each), using only core nodes; also includes a corrected file for the standard ComfyUI LoRA loader

Usage

  • ComfyUI: Place moss_sft_4step_rank_64_lora_comfyui.safetensors in models/loras with strength 1.0; use Comfy-Org/MOSS-Image's moss_image_sft_bf16 as the base model; use the euler sampler with Manual Sigmas "1.0, 0.91, 0.717, 0.024, 0.0" and CFG 1 (Basic Guider)
  • diffusers: Use the apply_4step_lora function in the card to attach moss_sft_4step_rank_64_lora.safetensors to the transformer, with num_inference_steps=4 and guidance_scale=1.0
  • Do not use ComfyUI's simple scheduler; keep strength at 1.0, as the author claims increasing it only causes double edges without adding detail

Risks

  • Only adapted for MOSS-Image SFT and its sampling schedule; sizes outside the training range, such as the default 2848×1600 in official templates, require very long prompts; large-size VAE decoding errors occur on Macs, requiring chunked decoding
  • Issues listed by the author: fabricated text in areas where text was not specified, similar crowds in vertical formats, inaccurate number of people at multi-person dining tables, occasional typos; comparison images are all generated by the author, base model license must be checked separately, and this entry has not been run

Open

Hugging Face Trending

Explore the latest AI Radar updates